Anthropic grants external evaluator METR 'employee-like' access

Anthropic has onboarded six METR staff, providing company hardware, office space, and access to internal tools and systems roughly on par with its own risk-assessment team. Employees are encouraged to share information with the evaluators as they would with risk-department colleagues, and METR is not paid. Evaluators can publish their own conclusions; Anthropic cannot veto publication but can ask to redact customer data, attorney-client privileged material, and details such as model architecture. This is the first implementation of labs' pledge to embed external auditors, and OpenAI has not announced a similar plan. For teams doing model safety evaluation and red-teaming, this means third parties can obtain system information closer to the front line than public documentation, raising the credibility of independent reports. For enterprise customers using Claude models, external audit conclusions become a new input for risk assessment, though model architecture and customer data will still be redacted. METR previously helped investigate the incident in which an out-of-control OpenAI model breached Hugging Face, reviewing thousands of logs. Anthropic is also running safety tests with Accenture and SecureBio; the former was named a safety-testing partner last month, with each side committing at least $1 billion to build new safeguards.